← All work
Working, not yet public

GitHub compliance evidence collector

Reads a repository's settings through the GitHub API and maps each check to a SOC 2 criterion, with the raw response saved as proof.

Python · httpx · GitHub REST API · pytest

  • 10 checksSource: Checks table in the project README
  • 3 outcomesSource: PASS, FAIL, UNKNOWN; engine.py
  • 8 of 8 tests passingSource: python -m pytest, run on 10 November 2026; tests replay recorded API responses

The problem

Before an audit, engineers prove controls like "no code reaches production without review" with screenshots and spreadsheets. The settings also drift between audits without anyone noticing.

What was built

  • 10 checks covering branch protection, review on recent commits, admin access, Dependabot, secret scanning, and security policy files.Source: Checks table in the project README; one collector function per check
  • PASS, FAIL or UNKNOWN comes from a deterministic rules table. No language model decides compliance.Source: rules.py and engine.py
  • UNKNOWN is not FAIL: a 404 on branch protection can mean "off" or "token cannot see it", so FAIL needs a second, unambiguous source.Source: engine.py; covered by the no-permission test fixture
  • Every API response is saved unmodified next to the result, and tests replay recorded responses with no network or token.Source: evidence/<owner>__<repo>/<timestamp>/raw and tests/fixtures

What this does not show

  • This demonstrates automated evidence collection. It is not a SOC 2 attestation, and the check-to-criterion mapping is my own.
  • It covers one repository. Organisation settings, audit logs and non-GitHub systems are out of scope.
  • The code is not on GitHub yet, so there is no link.

Receipts

ClaimWhere it comes from
10 checksChecks table in the project README
3 outcomesPASS, FAIL, UNKNOWN; engine.py
8 of 8 tests passingpython -m pytest, run on 10 November 2026; tests replay recorded API responses
10 checks covering branch protection, review on recent commits, admin access, Dependabot, secret scanning, and security policy files.Checks table in the project README; one collector function per check
PASS, FAIL or UNKNOWN comes from a deterministic rules table. No language model decides compliance.rules.py and engine.py
UNKNOWN is not FAIL: a 404 on branch protection can mean "off" or "token cannot see it", so FAIL needs a second, unambiguous source.engine.py; covered by the no-permission test fixture
Every API response is saved unmodified next to the result, and tests replay recorded responses with no network or token.evidence/<owner>__<repo>/<timestamp>/raw and tests/fixtures
NextKRISHIBODH: a gantry robot that farms a plant bed